{
 "model": "fusion-embedding-2-tactus",
 "version": "v0.1-preview",
 "benchmark": "STAG classification (Nature 2019), 27 objects, unseen test recordings, open-vocabulary text queries",
 "released_checkpoint_eval": {
  "top1_27way": 0.817,
  "top3_27way": 0.9507,
  "kway5": 0.948,
  "recording_top1": 0.7407
 },
 "recipe_mean_over_4_runs": {
  "top1_27way": 0.771,
  "top1_std": 0.062,
  "top3_27way": 0.935,
  "kway5": 0.935,
  "recording_top1": 0.722,
  "runs_top1": [
   0.8293,
   0.7014,
   0.7367,
   0.817
  ]
 },
 "baselines": {
  "scratch_mean_n3": 0.705,
  "stag_2019_supervised_cnn_top1": 0.76,
  "chance_top1": 0.037,
  "chance_top3": 0.1111,
  "chance_kway5": 0.2
 },
 "note": "Evaluation mirrors STAG's cluster-sampling test protocol, not their byte-identical harness. The recipe mean exceeds the 2019 CNN by less than one standard error; we describe the result as matching to exceeding the original baseline. Run-to-run standard deviation of 0.062 is a known limitation."
}